petrogurcak/cowork-setup

Cowork Setup Skill

Announce at start: "Používám cowork-setup skill pro [init shared layer | nastavení projektu <name>]."

When to Use

• User says "setup cowork project" / "nastav cowork projekt" / "cowork init"
• User starting non-development work in Claude Cowork desktop app
• User has dev projectsetup skill but it's overkill for content/marketing work

DON'T use this skill:

• Development projects (TDD, Git, build commands needed) → use development:projectsetup
• Adding workflow optimizations to existing setup → use workflow-optimization

Two Modes

| Mode | Trigger | Frequency | | ------------------ | ----------------------------- | --------------- | | A: Init shared | cowork-setup --init-shared | ONCE per user | | B: Per-project | cowork-setup <project-name> | Per new project |

If user runs Mode B but ~/Claude-shared/ doesn't exist → stop with: "Run cowork-setup --init-shared first."

Mode A: cowork-setup --init-shared

Run ONCE per user. Creates ~/Claude-shared/ shared context layer.

Pre-check

If ~/Claude-shared/ exists, prompt:

• A) Reinit — wipe and recreate everything (require explicit "yes, reinit")
• B) Update single file — pick which (about-me, voice-and-style, projects-index)
• C) Cancel (default)

Flow

1. Track intent marker. Before any mkdir, save list of paths skill will create. On user cancel mid-flow, delete ONLY those paths. Never touch pre-existing files.

2. Create directory structure:

   mkdir -p ~/Claude-shared/templates ~/Claude-shared/archive
   

3. Write working-rules.md — copy verbatim from templates/working-rules.md. No Q&A, no substitution.

4. Q&A for about-me.md (3 short questions):

   • Q1: "Jméno + role + kontext?" → <NAME>, <ROLE>
   • Q2: "Mission / co teď stavíš?" → <MISSION>
   • Q3: "Work blocks (kdy primárně pracuješ)?" → <WORK_BLOCKS>

   Optimization: Pokud existuje ~/.claude/CLAUDE.md se sekcí # ME, extract values + ask "Použít tyhle hodnoty?" → user confirm/edit.

   <TOOLS_CONTEXT> = optional, extract from ~/.claude/CLAUDE.md "Focus Areas" if present, else empty.

   Substitute placeholders in templates/about-me.md.template → write to ~/Claude-shared/about-me.md.

5. Voice extraction sub-procedure (see "Voice Extraction" section below):

   • Ask: "Máš samples textů (newsletter, IG, web copy)? [yes / no / skip]"
   • yes → run extraction
   • no → offer copywriting:brand-voice skill (Q&A mode, ~30 min)
   • skip → copy templates/voice-and-style.md.template to ~/Claude-shared/voice-and-style.md as-is

6. Initialize projects-index.md — copy verbatim from templates/projects-index.md.

7. Generate Cowork Global Instructions text. Print to terminal:

   I am <NAME>. When I open a Cowork project:
   1. Read ~/Claude-shared/projects-index.md, find current project, summarize: last date, what was done, next step.
   2. Follow rules in ~/Claude-shared/working-rules.md.
   3. Apply voice from ~/Claude-shared/voice-and-style.md.
   4. Never delete files without explicit confirmation.
   5. File naming: YYYY-MM-DD-slug.md.
   

   Followed by: "→ Open Cowork desktop app → Settings → Cowork → Global Instructions → paste this text."

8. Print UI linkage instructions:

   "When creating a new project in Cowork: Add Context → ~/Claude-shared/ AND your project folder."

9. GitHub bridge setup (optional, strongly recommended for claude.ai chat sync). Ask: "Init git + push to GitHub? [yes/no/later]"

   • yes: (a) gh auth status — if fail, instruct gh auth login and pause. (b) Ask repo path (default <gh-user>/claude-shared). (c) Write .gitignore from templates/gitignore.template. (d) Skip init if .git/ exists, else: cd ~/Claude-shared && git init && git branch -m main && git add . && git commit -m "init: shared context layer" && gh repo create <REPO> --private --source=. --push. (e) Append templates/working-rules-git-addendum.md.template to working-rules.md. (f) Print: "claude.ai → Settings → Connectors → Add GitHub → Authorize → Find <REPO>. Per-Project: Sources → GitHub → Select <REPO>."
   • no: skip, no git artifacts.
   • later: print manual instructions, skip exec.

Voice Extraction Sub-Procedure

Used in Mode A step 5 when user has writing samples. Grounds voice in real texts vs. Q&A.

Input

User provides 3-7 samples (file paths / URLs / inline text). Recommended: 2-4 newsletters, 3-5 IG captions, 1-2 web pages, 1-2 sales/product examples.

Extraction framework

Analyze all samples and extract 6 dimensions:

| Dimension | What to extract | | -------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------- | | Tone | Formality 1-5, energy 1-5, humor (none/light/heavy), expertise level | | Vocabulary | 10-20 preferred words, 5-10 banned words | | Sentence structure | Avg length, fragmentation yes/no, parataxis vs hypotaxis | | Brand codes | Recurring phrases (≥2 occurrences across samples) with example contexts | | Anti-AI patterns | Words/structures to avoid (em-dash overuse, "delight", "robust", "leverage") | | Per-channel adaptation | Differences between channels — ONLY if samples cover ≥2 channels. If 1 channel only, OMIT this section from output (do not write TBD placeholder). |

Output

Write to ~/Claude-shared/voice-and-style.md (NOT the empty template) with sections in order: Tone (formality/energy/humor/expertise) → Vocabulary (preferred/banned lists) → Sentence patterns (with 1-2 examples each) → Brand codes (recurring phrases with contexts) → Anti-AI patterns (what to avoid + why) → Per-channel (omit entirely if 1 channel) → Source samples (files/URLs analyzed with date).

Confirmation

Show draft to user → ask: "OK / change specific section / regenerate with different samples".

Edge cases

• No samples provided: fallback to copywriting:brand-voice skill (Q&A) or skip (write empty template).
• Inconsistent samples (extraction yields contradictory voice): ask "Want per-channel voice or unified? Add more samples from primary channel."
• WebFetch fail (URL unreachable): continue with remaining samples, notify user.

Mode B: cowork-setup <project-name>

Run for each new Cowork project. Creates ~/Projects/<project-name>/ and appends entry to ~/Claude-shared/projects-index.md.

Pre-check

• If ~/Claude-shared/ doesn't exist → stop: "Run cowork-setup --init-shared first."
• If ~/Projects/<project-name>/ exists → stop: "Already exists. Choose another name."

Flow

1. Q&A for about-project.md (4 questions):

   • Q1: "Goal — co chceš dosáhnout?" → <GOAL>
   • Q2: "Deadline?" → <DEADLINE>
   • Q3: "Deliverables — co produkujete? (list)" → <DELIVERABLES_LIST> (parse user list as bulleted markdown)
   • Q4: "Related projekty?" — read ~/Claude-shared/projects-index.md, suggest matches via string similarity in project name (split on -, match shared tokens). User confirms/edits/skips. → <RELATED_POINTERS>

2. Create structure:

   mkdir -p ~/Projects/<project-name>/OUTPUTS ~/Projects/<project-name>/REFERENCES
   

3. Write about-project.md — substitute placeholders in templates/about-project.md.template, write to ~/Projects/<project-name>/about-project.md.

4. Append entry to ~/Claude-shared/projects-index.md in ## Active section:

   FIRST_DELIVERABLE=$(echo "$DELIVERABLES_LIST" | head -1 | sed 's/^- //')
   ENTRY="- $(date +%Y-%m-%d) | <project-name> | started | next: $FIRST_DELIVERABLE"
   awk -v entry="$ENTRY" '/^## Active$/{print; print entry; next}1' ~/Claude-shared/projects-index.md > /tmp/idx.tmp && mv /tmp/idx.tmp ~/Claude-shared/projects-index.md
   

   No manual user step. Skill writes directly.

5. Print UI linkage instructions:

   "Open Cowork → New Project → Context → Add Folder:

   1. ~/Claude-shared/
   2. ~/Projects/<project-name>/ Set project name in UI: <project-name>"

Error Handling & Edge Cases

| Situation | Skill behavior | | -------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------ | | ~/Claude-shared/ exists, user runs --init-shared | 3-way prompt: Reinit / Update single file / Cancel (default) | | ~/Projects/<name>/ exists | Stop, suggest different name | | Mode B but ~/Claude-shared/ missing | Stop, instruct to run --init-shared first | | Voice extraction: no samples provided | Fallback to copywriting:brand-voice Q&A or skip | | Voice extraction: WebFetch fail | Continue with remaining samples, notify user | | projects-index.md append fail (file locked) | Retry once, else warn user + print line for manual paste | | User cancels mid-flow Q&A | Use intent marker — delete ONLY paths skill created in this session. Never touch pre-existing files. If no mkdir ran yet → no-op. | | Step 9: gh auth status fails / .git/ exists / gh repo create rejects | Auth fail → instruct gh auth login + retry. .git/ exists → skip init, verify remote. Repo exists → ask "Use existing?" + git remote add origin + push. |

Idempotence

• Mode A re-run after success → 3-way prompt (Reinit / Update / Cancel)
• Mode B with same <project-name> → stop, suggest different name
• All file writes are atomic (write to temp, then mv)

Out of Scope

• ❌ TDD/Git/verification rules (use development:projectsetup)
• ❌ Module catalog (brand-voice, audience wrappers) — Cowork calls skills directly
• ❌ Auto-sync script for projects-index.md — Cowork updates it via working-rules.md
• ❌ Claude Design setup — only documentation note (link REFERENCES/ from Design)
• ❌ claude.ai web chat integration — gap (no local folder linking from web)